Phase transitions of zirconia: Machine-learned force fields beyond density functional theory

Machine-learned force fields (MLFFs) are becoming an increasingly important tool in materials science and physics. However, most MLFFs are constructed based on density functional theory (DFT) calculations, which come with significant limitations. Here, the authors combine an efficient on-the-fly active learning procedure and a ∆-machine learning approach, enabling the generation of MLFFs with an accuracy that exceeds DFT accuracy at a modest computational cost. Using this method, they generated an MLFF for the random phase approximation that allows highly accurate predictions of the phase transition temperatures of zirconia.